Another perspective of temporal data mining is trend mining. Broadly speaking, trends are indicators of change over time with respect to some activity. The application of DM to time series data allows for the identification of trends and changes in trends. The research described in this thesis is concerned with the identification of temporal-spatial patterns (referred to as “combination patterns”) that change over a period of time. Reported work on trend mining has been directed at the forecasting of financial market trends based on numeric financial data, and the usage of text corpi in business news [118]. Other examples of mining trends using temporal and spatial data can be found in biomedical research [88], environmental protection [63] and road traffic management [112]. In this section, Sub-section 2.4.1 discusses previous work on trend mining, fol- lowed by Sub-section 2.4.2 which briefly discusses a few related studies concerning trend analysis.
Temporal-spatial Data Mining task
Descriptions Techniques
Static spatial data
Temporal-spatial data
Segmentation Clustering •Cluster Analysis •Temporal exten- sion to clustering Classification •Bayesian classifi-
cation •Temporal exten- sions to classifica- tion •Decision tree •Artificial neural networks Deviation and Outlier Analysis
Finding rules and re- lationships between at- tributes over time
•Association rules •Temporal associ- ation rules •Bayesian net- works •Temporal exten- sion to Bayesian networks
Trend Discovery Prediction of lines and curves, •Discovery of common trends Sequence mining Summarising temporal database, •Regression Discovering correlations
among the events in se- quences
Table 2.1: A possible classification of spatial temporal DM tasks and techniques [141]
2.4.1 Example of Types of Trend Mining
In the context of this thesis, trends are defined in terms of the frequency counts (sup- port) associated with individual patterns. Given this definition, it is possible to iden- tify some similarities with the work on Jumping and Emerging Pattern (JEP) mining. Jumping patterns are usually defined as patterns whose support changes dramatically from one time stamp to another. Emerging patterns are then a special form of jumping pattern where the support changes from belowσ to aboveσ over two consecutive time stamps. An example of work on JEP mining can be found in [67], where a moving window approach was adopted to identify JEPs. JEP mining has found application in a number of areas, one example is in medical research where JEPs have been used to monitor the progress of cancer cells [142].
In another study, found in [128], an iterative time-series trend mining mechanism was developed to identify associations in discovered frequent trends using categorical and continuous time-series datasets. The concept of frequent pattern trends defined in terms of sequences of frequency counts has also been adopted in [116] in the context of
longitudinal patient datasets. In [116] trends are categorised according to pre-defined prototypes and are grouped using a clustering method. The discovered trends from data episodes in this work were described as increasing (emerging), ups and downs, stableand jumping as shown in Figure 2.4.
Figure 2.4: Types of trends
Figure 2.5 shows an example of a time series of the form considered in this thesis. The time series given in Figure 2.5 is defined by the frequency counts for a patternX
collected over a period t. The granularity of t can be defined as required by the user (for example weeks, months, or years). Clearly from the given trend, useful information can be derived by observing where significant change points occur in the trend line.
2.4.2 Trend Analysis
Trend analysis refers to the concept of collecting information over a period of time to recognize positive or negative movement or changes in data (where the data typically describes an event or activity). While trend analysis is often used to forecast future behaviors, it may also be used to assess events in historical data. Trend analysis can be valuable as an advanced indicator of some potential problem or issue, for example decreasing trends in sales of a product line.
Trend analysis has been applied in many areas like health [64], climate [134] and human behavior [12]. Examples can also be found in the context of social network analysis. For example, Glooret al. [44] introduced a novel trend analysis algorithm to
0 5 10 15 20 25 30 35 40 45 50 1 2 3 4 5 6 Fr e q u e n cy oc cu rr e n ce s t Change points
Figure 2.5: Significant Change Points in a Trend.
generate trends from Web resources. The algorithm calculates the values of “temporal betweeness” of online social network node and link structures to observe and predict trends concerning the popularity of concepts and topics such as brands, movies and politicians. Likewise, some research directed at recommender systems [18, 143] and online market research [61] focuses on trends describing online social interactions and trusts so as to improve online marketing and sales strategies.
An important aspect of trend analysis is change point detection (some examples change points were highlighted in Figure 2.5). From the literature, a number of exam- ples mechanism directed at identifying change points can be identified. For example, Taylor introduced a control chart as a change point analysis tool to detect changes in data and describe the characteristic of these changes [122]. Change point detection has also been considered with respect to a variety of applications. In [40] a system for highlighting change points in historical temperature data was described. In [22] a system was described for detecting change points in Biodiversity measure trends so as to identify species habitat changes.